Skip to content

2.2 What the tools really do: rules, physics and search

The previous section listed the software in the workflow. This section asks what each tool is, underneath the interface — because the answer decides whether it counts as AI at all.

2.2.1 The honest-AI test

There is one question worth applying to every capability in this field, and we apply it throughout the study:

Does the system learn from data, or does it compute from equations?

Only the first is AI in any meaningful sense. The distinction matters because the moulded-parts software market is full of classical technology wearing AI labels — search algorithms sold as "autonomous optimization", rule engines sold as "intelligent" checks. The test cuts through the labels. It has nothing to do with how sophisticated or useful a tool is; a mould-flow solver is more sophisticated than most machine-learning models, and it is still not AI, because it computes the answer from physics rather than learning it from examples.

Applying the test across everything this field calls intelligent yields six mechanism classes, split cleanly in two. Three of them compute — the classical trio, the subject of this section, and the mechanisms behind essentially the whole toolset of §2.1. Three learn — the AI trio, and the subject of the next section.

computes

learns

Honest-AI test
learn from data,
or compute from equations?

Classical trio
R · Rules
P · Physics
O · Optimization

AI trio
K · Retrieval
ML · Machine learning
L · Language models

2.2.2 The classical trio: R, P, O

These are the three mechanisms behind almost every tool in §2.1. Each is genuinely useful; none of them learns.

  • R — Rules engine. A checklist a human wrote, applied automatically. The DFM checkers are the clearest case: "every wall parallel to the opening direction must have at least 1° of draft" is a rule, and the software just checks it against the geometry. Knowledge-based engineering (KBE) templates that generate routine CAD are the same idea. The knowledge is real, but it was put in by a person; the engine only applies it.
  • P — Physics solver. The mould-flow simulators and structural (FEA) tools. They take the geometry, the material properties and the process settings, and numerically solve the governing equations of flow, heat and shrinkage: equations in, fields out. Their proprietary value lies in validated solvers and measured material databases — thousands of lab-characterised plastic grades — not in anything learned from past parts. Run the same part twice and you get the same answer; the solver remembers nothing.
  • O — Optimization and DoE. Automated search. Topology optimization removes material from a design space over many repeated physics runs; a design of experiments (DoE) sweeps a grid of process settings and keeps the best result. This is the most-relabelled class in the market — it is what most "generative design" and "autonomous optimization" claims actually are. It is powerful, but it is search, not learning: it explores options using the rules and equations it is given, and it carries nothing from one job to the next.

2.2.3 The common thread — and its limit

Notice what R, P and O share. Every one of them computes from a model a human supplied — a rule, an equation, a search objective — and every one of them starts each new part from a blank slate. A solver that analysed ten thousand warped parts last year is no better at the ten-thousand-and-first; it re-derives everything from the physics each time.

That is a strength: the results are principled, repeatable and explainable. It is also the ceiling. These tools can tell you what the equations say will happen; they cannot notice that parts shaped a certain way tend to warp in ways the equations, at practical resolution, miss — because noticing tendencies across past examples is not something a rule or an equation does. Noticing patterns in accumulated data is exactly what the learning mechanisms are for, and the next section turns to them.